What is agentic AI?

Before diving into the risks, it’s important to understand how agentic AI actually works. At a high level, agentic AI tools are designed to take a task, break it down into steps and execute those steps autonomously. By iteratively reading and generating text, the tool aims to mimic reasoning and is able to call upon all kinds of tools to successfully execute its tasks, often using external tools such as APIs, databases or code execution environments.

The more access it has to data and tools, the more capable it becomes. However, these systems are not deterministic, they don’t follow fixed rules but instead make probabilistic decisions about what to do next. Now, the question of how well these tasks can be executed autonomously is food for another discussion. What is often overlooked, though, more autonomy and access also means that the impact of its mistakes, or manipulation by others, can be much bigger.

Where things go wrong

If an agent can act, it can also misact, and this can happen in different ways. Let’s look at 3 concrete risks.

1. Dataset poisoning: With dataset poisoning, the foundational model used as the brain of your agent is already being attacked before it has been trained. Because the model relies on vast amounts of internet data, it’s often collected at scale with limited filtering. The internet is riddled with examples of how things should not be done: in software development, for example, lots of open source codebases are filled with vulnerabilities. But what if a hacker utilizes this to specifically target you? Research funded by Anthropic investigated just that and found that as little as 0.00016% of the training set has to be directly poisoned to be able to introduce a backdoor. These backdoors can be triggered by specific phrases, causing the model to behave in ways defined by the attacker
2. Prompt injection: A more direct and immediate threat is prompt injection. Here, the attacker doesn’t target the model itself, but its input. By manipulating what the model reads, they can influence how it behaves. An attacker could host a legitimate website containing hidden instructions, text invisible to the user but readable by the model. These instructions tell the agent to extract sensitive information (account credentials, banking information, API keys, …) and send it elsewhere. Even with clear safeguards in place, researchers pulled this off using a surprisingly simple approach, making it clear that it’s far easier to exploit than you’d expect.
3. Slopsquatting: Even without direct manipulation, agents can create risk through their own mistakes: They have an inherent tendency to hallucinate. Since LLMs probabilistically produce text (what word would be most probable to follow in the sentence), they do make mistakes and present things that don’t exist. Attackers can exploit this. If an agent tends to invent a dependency name, an attacker can register it and attach malicious code. When someone else’s agent later tries to install it, the system becomes compromised. This technique has been dubbed as slopsquatting, as the AI variant of typosquatting, and it turns model errors into attack vectors.

How to mitigate the risks

All of these vulnerabilities stem from the same underlying issue: the agent can act autonomously. In theory, the safest approach would be to manually approve every action and verify every output. But doing so removes the very benefit agents provide: autonomy. So what are your other options?

Sandbox the agent

Agents should never run directly on your core systems, production servers or sensitive data with unrestricted access. How can you implement it:

• Run the agent in a container, virtual machine or isolated cloud environment.
• Define and limit access to only the directories, files or databases it actually needs for its tasks.
• Prevent it from installing arbitrary packages or executing commands outside its environment.

If something goes wrong, whether due to a model hallucination, a prompt injection attack or an external exploit, the damage is confined to the sandbox, protecting your main systems and customer data. There are also specialized sandbox processes that can wrap around your agent and do this confinement for you.

Control external communication

Limit which websites, APIs or network endpoints the agent can access. You implement it as follows:

• Use egress filtering with a deny-by-default configuration, only allowing connections to trusted domains.
• Require human approval for each request to an untrusted domain.
• Monitor outbound requests to detect unexpected activity.

This reduces exposure to prompt injection attacks or slopsquatting, where malicious content on a web page could trick the agent into revealing secrets or executing harmful commands.

Apply least-privilege access

Give the agent only the credentials, permissions and data it absolutely needs for its task. How to implement it?

• Use user-scoped tokens rather than global or admin credentials.
• Make tokens time-limited so access expires automatically.
• Restrict permissions to the minimum required actions (read-only where possible).

If an agent is compromised, leaked credentials or tokens won’t grant full access to sensitive systems, reducing potential damage.

Keep humans in the loop

Agents can act autonomously, but critical actions should always have oversight. You implement it as follows:

• Require review for any actions affecting production systems, customer data, or financial transactions.
• Log agent decisions and outputs for auditing.
• Where possible, set up a system that allows agent actions to be rolled back.

Humans can catch mistakes or malicious behavior that the agent might miss, while still letting the agent handle low-risk tasks to improve efficiency.

Final thoughts

Agentic AI is powerful precisely because it can act independently. But that same independence creates a new attack surface. Not just when using a (free) agentic AI tool online, but also when building a custom tool for your business. Want to learn more about how we use and build AI tools responsibly? Make sure to reach out to our AI specialists!

The post How to secure agentic AI in software development appeared first on Verhaert Masters in Innovation.

The post How to secure agentic AI in software development appeared first on Verhaert Masters in Innovation.

Base imagery © 2023 Hexagon and data partners

A rising demand for high-quality imaging

In today’s imaging industry, the demand for ever-higher image quality is relentless, yet constraints on cost, weight and size continue to challenge engineers. Meeting these requirements often means more complex or larger hardware, pricier components, and increasingly demanding testing, all of which push projects toward higher budgets and longer development cycles.

Meanwhile, AI-driven image enhancement tools are rapidly advancing and becoming more widespread, from image deblurring features in Lightroom and Photoshop to AI upscalers and generative models like Stable Diffusion. This raises a compelling question: Can artificial intelligence remove blur and improve usable image resolution without altering the existing optical hardware? In other words, could we now finally achieve the kind of dramatic ‘image enhancement’ often depicted in popular TV shows in real-world applications?

Rewriting image quality with AI

The answer? Kind of. The big breakthrough is that artificial intelligence can now act as a kind of ‘virtual upgrade’ for optical systems. By training models on images from high-end cameras or sensors, developers can teach AI to reconstruct what a better system would have seen. In practice, this means cheaper, smaller and lighter optics can produce images that look like they came from far more expensive hardware, dramatically reducing overall system costs.

There are two especially promising approaches. The first is to use AI to boost the image quality of cheaper cameras and sensors, allowing low-cost, lightweight hardware to produce results closer to premium systems. This can also be used in constellations of sensors or satellites: a few units carry premium optics, while many others use simpler hardware. The second is retrospective enhancement: applying AI for image deblurring in hardware already deployed in the field, improving quality without any physical upgrades. Together, these approaches show how AI can extend the capabilities of both future and existing optical systems, purely through software.

At its core, this technique is generative: the AI starts with a degraded image and generates a sharper, higher-quality version. This only works well if the distortions and noise of the real optical system are accurately modeled during training, which requires deep technical expertise. A strong starting model is also crucial, using an AI already trained to deblur regular images makes adaptation faster, cheaper and more reliable. This is how you can reuse and fine-tune existing models to turn image enhancement from a research challenge into a practical engineering tool.

Base imagery © 2023 Hexagon and data partners

AI in action: From space to surgery

AI-powered image enhancement is starting to improve real-world systems across the board. Any product that relies on optics can benefit: from drones monitoring traffic and satellites mapping crops and deforestation changes, to smartphones capturing sharper photos without bigger cameras. Even weather alerts and environmental policy enforcement can be more accurate thanks to clearer, AI-enhanced imagery.

In optical systems for life sciences, AI image deblurring can boost standard microscopes by leveraging training on high-resolution systems. This approach offers high-resolution capabilities to millions of conventional microscopes in the field, improving live-cell imaging, digital pathology and microbial detection. It also supports low-cost diagnostic devices and high-throughput drug screening. This will help researchers and clinicians gain clearer, more reliable insights without upgrading hardware, making advanced imaging more accessible, scalable and cost-effective.

Generative constraints and reliability

AI-based image deblurring is powerful, but it isn’t magic. It can’t recover information that was completely lost in the original image. Since it’s a generative process, it can also introduce artifacts. This is inevitable, but in most cases, not a dealbreaker. For example, imagine a blurred photo of a pedestrian holding a small object. After deblurring, the pedestrian becomes clearer, but the artificial intelligence can only make an educated guess about the object. They might reconstruct it as a phone, a coffee cup, or something else entirely, but for tasks like counting pedestrians or detecting presence, this isn’t a problem.

The AI works much like humans do: it interprets shapes and context to reconstruct likely features, but the accuracy depends heavily on the quality and variety of training data. That’s why high-quality images across diverse conditions and blur types are crucial. By understanding these limits, AI deblurring can be used responsibly, enhancing images without creating misleading or false information.

A new era of optics and AI

AI-driven image enhancement is already proving to change the game tremendously. It can boost the performance of cheaper cameras and sensors, and even improve images from hardware that’s already out in the field. Better-quality imagery at lower cost could transform Earth observation, from tracking traffic and deforestation to responding faster to disasters. While the impact will be huge, broad adoption will likely happen step by step. As artificial intelligence continues to get smarter, the line between what optics alone can achieve and what software can deliver will keep blurring, opening new possibilities for how we capture and use images across industries.

The post Blurring the line between optical hardware and AI appeared first on Verhaert Masters in Innovation.

The post Blurring the line between optical hardware and AI appeared first on Verhaert Masters in Innovation.

Why high-quality data is essential for valuable AI

AI models learn patterns and make decisions based on the data they’re trained on. If the input data is flawed, the output will be unreliable: garbage in, garbage out. Because AI systems require substantial time and effort to build, you’ll want certain assurances it’ll be worth the investment. That’s why high-quality data is essential for ensuring accurate predictions, valuable insights, and effective automation.

And why poor data quality allows for serious risks. Bias in training data can result in unfair or misleading outcomes. At the same time, inconsistencies and missing values can cause AI to misinterpret information, leading to unpredictable results. And outdated data may drive decisions based on obsolete trends. The biggest threat? AI might generate results that appear correct but are fundamentally inaccurate, making it difficult to distinguish between right and wrong. Without a strong data foundation, even the most advanced AI models can fail. So when do you have a strong data foundation?

What about using AI as a data cleaner?

AI isn’t just dependent on high-quality data, it can also help improve it. Intelligent algorithms can fill in missing values for more complete data. Anomaly detection helps identify outliers and inconsistencies that might otherwise go unnoticed. AI can also identify redundant or conflicting entries, streamlining datasets. Finally, normalization and standardization enforce consistency, making data more structured and reliable for AI applications.

However, AI-driven data cleaning isn’t foolproof. If there’s a flaw somewhere in the process, AI may reinforce incorrect patterns rather than fix them. To prevent this, organizations should implement safeguards: either maintain a small, verified dataset for validation or have human experts review AI-refined data. Even with these precautions, AI-driven data cleaning can significantly speed up the process.

Fixing data at the source

Although AI-driven data cleaning can be a powerful tool when you know you don’t have high-quality data, there are other, sometimes better, options.

• Generating synthetic data: Simulations can be a fast solution for data scarcity. By defining specific parameters, simulations can produce endless data, helping identify trends and patterns when real-world data is lacking or insufficient.
• Leveraging external data sources: Both paid and freely available, external data from sources like governments, research institutions, and other companies can enrich your dataset, providing additional context and improving the AI model’s accuracy.
• Using alternative data sources: Gathering data from different systems or creating new collection methods can improve data relevance and accuracy. For example, you might not have a sensor in place that captures when a machine breaks down, but maybe your conveyor belt stops working and sends an alert as a result, which can be captured.
• Improving existing databases and collection methods: A more long-term solution is focusing on improving their existing databases and data collection methods. Even when you’re still building the AI system, better-designed data systems can automate reporting and even simulate basic insights AI might provide in the future.
• Implementing federated learning models: In privacy-sensitive sectors, federated learning allows you to train AI models using decentralized data without compromising privacy, as the raw data is never shared across systems.

From imperfect to powerful

High-quality data is the cornerstone of effective AI, but no dataset is perfect. By leveraging a combination of continuous improvements and proactive data-gathering strategies, businesses can refine their data to unlock their full AI potential. So how do you know which strategy will work best for you? You have to define what you want to do before you define how to get there. Typically, this is why we set up dedicated ideation workshops in which we get to the bottom of why you want to invest in AI and what you want to get out of it. This helps define clear data strategies, assess data quality, set up improved data-gathering processes, or even algorithmically improve your existing dataset. That’s how, with the right approach, AI can transform not only your insights but also the way you manage and improve your data over time.

The post Garbage in, garbage out: The importance of data quality for AI and how it can be improved appeared first on Verhaert Masters in Innovation.

The post Garbage in, garbage out: The importance of data quality for AI and how it can be improved appeared first on Verhaert Masters in Innovation.

In this blog, we’ll do a little deep dive into the different types of robotic and artificial intelligence systems that are already being developed. Let’s start with the current state-of-the-art certified robotic systems present in surgery rooms!

How can robots contribute to better healthcare?

Competent medical staff is invaluable for the healthcare system, but specialists are incredibly rare. The expertise scarcity means that a large portion of resources (time, money, fuel) is being spent on transport instead of performing medical procedures (regardless of if the patient or the doctor needs to travel). This is where teleoperated robots can get handy, as only the information (the robot’s real-time motion control) has to travel instead of people physically.

Not to mention that training someone to become a surgeon takes many years and many more years are needed to become a specialist. A computer-based system can be ‘trained’ almost instantly by updating its software: not in the same meaning as training a machine learning model, which can be an extensive process, but with the meaning of deploying an existing model for example. Of course, first, we need the mechanical capabilities and the software stack. That’s something we’re already working on.

Moreover, what are robots best at? Repeating motions over and over again with superhuman precision. Repeating motions over and over again with superhuman precision? Yes, repeating motions over and over again with superhuman precision (author’s attempt at humor). Therefore, the execution mistakes due to ‘human error’ could be significantly reduced. This discussion can be layered and tackled in detail:

• Robots as extensions: teleoperation with full human guidance and control.
• Robots with limited critical autonomy: teleoperation with partial human guidance and control.
• Robots with high critical autonomy: minimal human intervention during the medical procedure.
• Robots with autonomy on non-critical tasks: guiding, servicing and rehabilitation.
• Artificial intelligence enhanced healthcare: from perception and planning for robotics to volumetric registration for medical imaging and biosensors data analysis.

All these systems aim to:

• increase global accessibility of specialized medical expertise
• provide an early potential diagnostic
• optimize time and performance of medical procedures
• improve treatment/rehab consistency by making it fun

Da Vinci Surgical System

Robots as extensions

As mentioned in the introduction, we will discuss the state-of-the-art systems in terms of certified surgical robots. The most popular surgical robot is probably the ‘Da Vinci Surgical System’ (picture above). Although it represents a consistent technological advancement, the system isn’t conceptually complicated overall. In this setup, the robot is an augmentation of the medical practitioner via an optical system and a tactile interface. Basically a screen and a joystick controller (using simple words to describe a rather complex and sophisticated device).

Most of the surgical robots today fall in this class as they’re capable of giving superhuman perception and motion precision abilities to doctors. The teleoperation system acts as a scaling mechanism that can greatly improve the motion accuracy: “if the operator moves a joystick 10 cm, while the surgical robot moves an instrument only 1 cm, then the precision is 10 times increased. And it can be even much more. The optical system offers a focused and immersive connection between the operator and the surgical procedure and it can be further enhanced by augmented reality.

The drawback of these teleoperated surgical robots is that most of them do not have haptic feedback, meaning that the operator doesn’t ‘feel’ resistance when touching the tissue. There’s no force feedback control and therefore the responsibility (and the merit of success) is entirely attributed to the surgeon operating the system. Although they don’t possess substantial autonomy, these medical robots are highly robust and are currently the only type of medical device certified to be used in actual medical procedures.

While being a rather sensitive (and highly confidential) topic, we can proudly say that we have acquired some experience in building demonstrators of robots with autonomy on non-critical tasks, robots with limited critical autonomy, robots with high critical autonomy as well as AI-enhanced healthcare use cases.

Several types of robotics and AI systems are currently already present in healthcare, certified and capable of enhancing surgeons’ skills. However, many technologies and advances are popping up as we speak. In our upcoming blogs, we’ll tackle more types of robotic and artificial intelligence applications in the healthcare field.

Nevertheless, if your curiosity is greater than your patience, don’t hesitate to contact us and pick our brains on what we can do for your medical application!

The post From science fiction to reality: AI & robotics in healthcare appeared first on Verhaert Masters in Innovation.

The post From science fiction to reality: AI & robotics in healthcare appeared first on Verhaert Masters in Innovation.

Let’s start by clearing out two main misconceptions about the Robot Operating System (ROS). First of all, it’s not an operating system (OS) and it’s not just useful for robots. But since a ‘Not-Only-For-Robots Not-An-Operating-System’ isn’t quite as catchy, let’s refer to it by its original community-accepted name. Secondly, in this blog we’ll only refer to the robotic mechanical devices and NOT to robotic process automation software.

ROS, what?

What is ROS then? It’s actually a middleware. This means that it’s supposed to be placed between software logic and hardware. It’s similar to a real operating system but needs an operating system platform to run on (usually a Linux distribution). It can’t simply run directly on basic hardware. Moreover, devices need an ROS-specific interface to make use of their power, similar to how operating systems need drivers to access hardware resources.

So if it’s not just for robotics, why does it have ‘robot’ in the name? Because it was developed by the robotics community to be used with robots. Even though it can be used with applications that are not robots, most of the development efforts are directed toward perception, navigation and manipulation tasks. This means that highly complex robotic functionalities are already available in the ROS ecosystem.

ROS, how?

Let’s discuss the core power of ROS: nodes and topics. Nodes are somewhat the atomic pieces of software that implement a desired logic, like retrieving sensor data, capturing images, processing input data or sending input to an actuator. Topics are the communication links between nodes, independent entities that nodes can write or read data from – imagine them as lists of messages like internet chat rooms. These simple mechanisms allow the creation of complex distributed and robust architectures, where the overall system isn’t entirely dependent on the individual. As long as the system design allows it, the entire system is sensor and actuator-agnostic. Imagine a random vacuum cleaner robot: the navigation module doesn’t care if the localization uses lidar, ultrasonic or camera data, or if it’s using wheels, tracks or a propeller to move around.

These components should be easily replaceable, the only necessity is the right ROS node to take the hardware input and offer a standardized output on a specific topic (a driver-like piece of software, as we mentioned before). And here’s another very cool thing: each node can be on a separate computing platform or as many nodes as possible can share the same computing platform at the same time. It’s all up to the application architect to decide. The only prerequisite for ROS is to have a special ‘master’ node running that can be accessed by every other node. The drawback is that if the computing platform running the ‘master node’ disconnects from the system, then the entire system fails.

ROS topology example and in an autonomous car application ©VERHAERT

Above you can find an example of such a topology, as well as applied to a simplified autonomous car application. Keep in mind that each one of these nodes can be on the same or different computer as any of the other nodes. You can even deploy this system on multiple autonomous vehicles and then simply add topics that connect multiple vehicles. The limit of the scalability is set only by the computing or networking resources available.

ROS, why?

‘Standardized’ is another keyword for ROS. Its main goal is to offer a base for true rapid robot prototyping. In 2006, Eric and Keeman decided to create a standard framework for all robots (not yet ROS) when they noticed that in software projects there’s little time left to innovate because most of your time is spent simply making the sensors or actuators work with your system or setting up a communication architecture. The ROS is the successor to their initial framework and it fills the shoes perfectly. No more reinventing the wheel. Once a sensor, actuator or algorithm is successfully interfaced by a developer (or a group), it gets an ROS package made available for everyone to use!

Source: Eric and Keenan pitch deck

We’ve mentioned ‘true rapid prototyping’ before, which might have triggered your spider sense if you’ve spent some time between demonstrators, prototypes and products. Some people tend to oversell the former as the latter, but that’s not the case here. ROS is indeed a prototyping system: it’s not robust enough, doesn’t have real-time capabilities and is easily hackable. Moreover, the master node is required to function at all times.

ROS, next?

ROS2, on the other hand, has a completely different low-level implementation of ROS. It’s still based on nodes and topics, but is far from being a prototyping project and aims to be an industry-ready fully-distributed framework. It has more complex robotic libraries for navigation and manipulation that implement state-of-the-art algorithms and is designed so you can’t develop your packages without abiding by proper software engineering rules. It’s now harder to just create ‘something dirty that works for now’.

Other major changes include the ability to encrypt topics and time synchronization is being handled by the Data Distribution Service, a software framework to reliably and securely communicate in real-time. Furthermore, ROS2 doesn’t have a master node, meaning the application can be entirely distributed across multiple devices without having the inherent critical component. Pretty neat I would say.

Let’s reiterate the key aspects of ROS:

1. Core simplicity
2. Standardization of components
3. Native compatibility with distributed systems
4. Industrial readiness (on the way with ROS2)
5. Available for embedded systems (on the way with ROS2)

It’s clear that ROS is a software to watch. Whether you’re developing a fleet of drones, robot vacuum cleaner, robotic manipulation application, coffee machine or smart beer dispenser, make sure to add this awesome open-source framework in your trade-offs!

Hungry for more? Check out our other blog posts to see how ROS connects hardware with AI!

The post Robotics reinvented: a beginner’s guide to ROS appeared first on Verhaert Masters in Innovation.

The post Robotics reinvented: a beginner’s guide to ROS appeared first on Verhaert Masters in Innovation.

Trends in NewSpace

What is going on that makes earth observation suddenly so accessible? In NewSpace we see that the satellites have become relatively small. Decreasing the material cost and the cost of bringing them into orbit.  This makes it easier for companies to have a presence in space. There has also been a shift from single expensive satellites to constellations of satellites. These consist of duplicates of a satellite that are spread across the globe to increase coverage and reduce costs. Think of a famous billionaire whose space company is doing just that.

In short, companies can now also acquire earth observation data. Aside from using data themselves, they also gain revenue by selling it. This trend of earth observation data being commercially available has truly made earth observation accessible to everyone. By paying a subscription fee, anyone could acquire a monthly quote of data. There even exist platforms where you can get earth observation for free! You are certainly familiar with Google Earth, one such provider of free earth observation data. As you can see, acquiring data has become quite easy. But what makes it so valuable?

What can I do with earth observation data?

The most important thing you can do with this data is to automate the generation of insights. It doesn’t matter if you are a shipping company managing a fleet, an energy company forecasting energy demands, or a farmer wanting a continuous overview of all your crops. You won’t be doing this manually. As the scale of these tasks has become so large, you need to automate this process as much as you can. This is where AI comes in.

As the world gets more globalized, and all our human endeavors scale, automating the generation of insights becomes mandatory in order to not fall behind.

A maritime shipping company might gain insights into how busy certain routes are, how the weather and sea are behaving, or what the ice formations in the arctic look like. A farmer could gain insights into the status of crops. Think not only of the health and readiness to harvest, but the yields as well. To give a few examples you might already have an idea for an application that will generate a lot of value for you. But first of all you will need data, and not just any data, you need the right data. What kind of sensors do you need? You have a lot of options to choose from, from selecting specific spectral bands to LiDAR and other types of sensors. For each of these, you can also determine their characteristics. Then you probably have some requirements for the resolution which can range from a couple of hundred meters per pixel to an incredible thirty centimeters per pixel. Another important requirement is the fly-over time, which determines how long it takes for a satellite to pass by the same location twice. And of course, we cannot forget that there is a cost to all this. Premium EO data can go from a few cents to over 50 euros per square kilometer.

Spring cleaning

Before you can start building applications, you need to do a bit of cleaning to make the data ready for use. Depending on the provider of your data this can be a lot of work or none at all. We distinguish three main steps that typically have to be taken. First is noise removal. Noise sources like stray light, smear and dark offset can automatically be dealt with by using specialized algorithms. Second is pixel labeling, meaning that for each pixel in every picture the location and timestamp need to be determined. Luckily, AI can help you in this early stage as well. There are plenty of algorithms available for this purpose.

The third step can be quite tricky. If you have selected multiple sensors to measure the same phenomena, you need to correct for inter-instrument bias. What do we mean by this? Each of these sensors should be measuring the exact same thing. However, slight differences might occur resulting in different values for the same phenomena. Incorrectly using this raw data could be disastrous for your project. It is of vital importance to correct this bias first! Fortunately, machine learning algorithms can be employed to this bias and correct it, even if we don’t fully understand where the bias is coming from.

Supervised vs. unsupervised model ©VERHAERT

Applications

Spring cleaning is done and the data is now ready to use. It’s time to actually build algorithms that can extract insights from the data. Typically, you would want to use a supervised learning algorithm for this kind of problem. But, that means we need examples to teach the algorithm what it should do. And I don’t just mean a few. A large dataset needs to be gathered and manually labeled in order to be able to build a good algorithm. You want to determine the air quality from spectral imaging? Well, then you need to make a lot of manual measurements first so the algorithm can learn from your examples.

Of course, this kind of approach is very expensive, so don’t we know a better way? Within the world of machine learning there also exists a group of unsupervised learning algorithms. These don’t need examples at all. One of the more interesting approaches to earth observation goes as follows:

1. 1. You select the sensors which should capture your problem quite well
   2. You apply a clustering algorithm with a high number of classes on this data. This algorithm can divide the world into similar parts based on this data.
   3. Finally, it’s up to a domain expert to give meaning to these classes.

Let’s go back to the air quality example. If the right information is contained in the data, the clustering algorithm will divide the world into parts with similar air quality. It is then up to a domain expert to give meaning to these classes in order to make the algorithm usable. Measurements still need to be made, but far less than before. The result is an approach that is much cheaper both in time and resources.

Conclusion

Leveraging AI techniques enables us to automatically generate insights from the vast amount of earth observation data that is at our disposal. In these times of rapid globalization, this is a must as we just can’t keep up performing these tasks manually. If you want to generate insights from the vast amount of earth observation data that is at your disposal. And even if you want to generate these insights manually, you can benefit from AI to clean up your data.  No matter how you look at it, AI has become critical in any earth observation project we can imagine. Applying it unlocks a whole new area of opportunities.

Want to learn more about AI in data observation? Take a look at this webinar session!

The post AI, the rising star in earth observation appeared first on Verhaert Masters in Innovation.

The post AI, the rising star in earth observation appeared first on Verhaert Masters in Innovation.

What does it take to start considering and implementing resource-aware AI systems? Prioritizing tasks dynamically and intelligently based on computing resources is the way to go, especially together with embedded platforms that are meticulously energy efficient. Find out how they can satisfy the desire to significantly reduce AI energy consumption during this webinar from Niels Verleysen, Consultant AILab, and Bart De Vos, Consultant EmbeddedLab.

Watch the recording

View the presentation

The post The future of energy-efficient AI systems appeared first on Verhaert Masters in Innovation.

The post The future of energy-efficient AI systems appeared first on Verhaert Masters in Innovation.

Research and innovation at GITEC and AB InBev are always in motion. How are they coping with the current crises? What hurdles are they facing and how do they find new opportunities? Find out from Simon Gerdesmann, Global Director Digital & Open Innovation at AB InBev, how they are using AI and digitized processes to streamline innovation research, knowledge sharing and resources.

Watch the recording

The post Applying digital tools to advance innovation research appeared first on Verhaert Masters in Innovation.

The post Applying digital tools to advance innovation research appeared first on Verhaert Masters in Innovation.

Different approaches

Let’s start with the most basic approach, reactive maintenance, in which you only react and carry out maintenance when something breaks. This approach maximizes the usage of a machine but can result in unforeseen breakdowns. In a production line reactive maintenance typically isn’t preferred, as an unforeseen breakdown can bring production to a halt.

An approach that deals with this is preventive maintenance, in which you’ll carry out maintenance on a schedule. The idea is to ensure nominal machine performance by doing maintenance before machines break. The great benefit here is that you can schedule when your machine will be unusable. The cost, however, is that you are performing maintenance early.

If you remember the title, you already know where this is going. To prevent the added cost of preventive maintenance, you can use machine learning to give insights in the state of the machine. Based on these insights, the predictive maintenance approach allows you to schedule maintenance only when it’s necessary.

Possible machine insights

There are many different ways to gain useful information for maintenance. One approach is based on failure classifications. Depending on how you frame this task, you can not only predict if a machine will fail within a time frame but also what type of failure it’ll be and what caused it.

But what if you want to have an exact prediction of a machine that could break down? Look no further. Failure regression allows you to predict when the machine will most likely fail and know well in advance when maintenance is necessary or how long you can postpone halting the machine.

Not all failures are predictable though. What if someone uses a machine the wrong way or something hits the machine in a vulnerable place? For this, you need to somehow know if your machine is working correctly right now. This task can be done through anomaly detection algorithms which will tell you at any moment if a machine is still working correctly or not.

A final category is survival analysis, which models the machine’s degradation process. This way you can gain insights in how the machine degrades and which activities or environments make it degrade faster or slower. This approach allows you to plan maintenance depending on the degradation level, as well as learn how to optimally use the machine. And by doing so, you increase the time between maintenance and make the machine more profitable.

Predictive maintenance requirements

Predictive maintenance all starts with data and the infrastructure to get this data. Any relevant data that indicates the correct functioning is interesting, including generic information from the machine, the gauges and other sensory equipment that is already available. People can see, hear or even feel if something is wrong or if the machine requires maintenance, like rust, creaking, or vibrations. These are all things that can be measured either directly or through virtual sensors, for instance by placing cameras with algorithms to detect rust or microphones for sounds and vibrations. This data corresponds to the state of the machine.

To learn how machines degrade, you also need to know how they have been used. Through work order and inventory usage data, for example, algorithms can learn that a period of heavy usage has a greater effect on the degradation than the holidays. More often than not, data from your CRM and ERP systems contain more useful information than the sensory equipment on the machines themselves, especially if you combine the data of all identical machines. That way you can reuse the same models on multiple machines and acquire more data to train and improve your models.

So how will you collect all that data? A good IoT infrastructure is going to be crucial. Installing such an IoT system would involve placing additional sensors, cameras and microphones, among others, and then connecting these to a network. Through this network, the data from these devices can be collected, cleaned and stored in a server, and then be used to create algorithms for predictive maintenance. Even if you don’t necessarily need or want predictive maintenance, such an infrastructure would be a great investment because the benefits go much further, like for data analyses or enhancing the work of monitoring teams.

Although the investment in the right infrastructure can be rather large, the potential benefits of predictive maintenance can be even larger. You can boost your overall equipment effectiveness. You can increase availability by reducing unplanned and planned stops. You can monitor the performance of your machines more closely and increase the overall machine performance quality. All while reducing the number of resources used and waste produced, bringing you closer to lean manufacturing.

Of course, an easily replaceable tool isn’t worth the investment of a predictive maintenance algorithm. So where do you start? By determining what kind of predictive maintenance your business could benefit from. If you are looking for guidance in this innovation journey, check out our AI services or get in touch!

The post Predictive maintenance – know when things will break appeared first on Verhaert Masters in Innovation.

The post Predictive maintenance – know when things will break appeared first on Verhaert Masters in Innovation.

Collecting datasets

In our previous blog, we defined different steps to discover your AI opportunities. Once you’ve identified the process you want to automate and the information you hope to obtain, you’ll need data to feed the model. These are the camera images, audio signals, text messages, or sensor measurements the model will analyze to provide you with answers to your questions. Whether you are looking to predict the stock market or develop a medical application, having low-quality, biased or unreliable data makes the task impossible. Take for example a study on blood oxygenation levels that fails to consider the difference in sensor response of the pulse oximeter between patients with different skin colors. This would significantly reduce the probability of detecting occult hypoxemia in black patients compared to white patients.

Problem understanding is indispensable to producing a valuable dataset. Your team should understand the variability relevant to defining the problem in practice. Often, people tend to overly bias the dataset toward the most accessible data. A self-driving car whose algorithms are trained only on roads the developers happen to travel regularly is not robust. Not entirely unlike humans, ML algorithms might find it challenging to assess unknown situations. For Machine Learning models, this results in unpredictable model outcomes because machine learning models can’t learn outside the data. So high volumes of information gathered in various circumstances are crucial to developing a trustworthy algorithm.

Finetuning the labeling process

The importance of data as a crucial building block in a machine learning project is gaining recognition. However, apart from raw, high-quality data, a machine learning project is built upon the data labels. They’re the ground truth of your model and represent the outcome your model should output. Think of it like this, a parent won’t just point at items to show their baby, they will also say the name of the item. This way the baby will learn to recognize and name these items in its surroundings. With a machine learning algorithm, this is exactly the same.

Obtaining labels can be complex and labor-intensive. Machine vision problems often require manual labeling for specific objects in each image. Depending on the application, the human labelers must have the appropriate qualifications to label medical scans, images of technical defects or any other specific image type.

Some things to consider during the labeling process:

• Different labeling requirements come at different prices. Only requesting a classification label for the complete image is a tenth of the cost of delineating all instances in the picture. The figure below illustrates different labeling approaches.
• While developing the model, it pays off to evaluate the current weaknesses so you know which labels you need to improve. Knowing what the model struggles with allows you to maximize the return on new data.
• When you outsource the labeling task to specialized companies, these are critical suppliers. Your team should treat them as such. You should monitor their results adequately. Too often, the perceived simplicity of the task makes people forget to define strict and well-thought-out quality metrics on the results.

Illustration of different label types. Point annotation (top left) costs less than full mask labelling (bottom right).
Squiggles (top right) and bounding box annotation are in between these extremes. (Source)

Maximizing the return

A dataset of delineated images is necessary to build a model to delineate objects. Currently, techniques are being developed to train models based on weakly supervised data. These techniques aim to use latent information in cheaper, low-information labels to prepare models for high-information output. In the classical approach, models require the same level of information in the labels and the desired result. This is expensive, so you’ll need a human to provide you with ‘examples’ of this valuable output.

Whether you are building an algorithm to read text documents or you are building a self-driving car, the message is the same. You don’t just need data, you need a high volume of qualitative data in all relevant circumstances and you should definitely not forget to gather qualitative labels. Do this and you’ll be one step closer to the optimal solution for your next ML project. Interested in learning more? Subscribe to our AI blog mail or visit the AILab page.

This article was co-written by Jan Alexander.

The post ML’s elephant in the room: data labeling appeared first on Verhaert Masters in Innovation.

The post ML’s elephant in the room: data labeling appeared first on Verhaert Masters in Innovation.